Respiratory motion represents a significant source of uncertainty in radiation therapy,
especially for lung tumors. The displacement of the tumor is affected by the biomechan-
ics of the lung, by anatomical variability, and by breathing patterns that cannot be reproduced.
The traditional motion models of the respiratiry system assume that respiration is periodic or
repeatable, which does not entirely capture the realities of clinical breathing patterns. In clinical
settings, the amplitude, frequency, baseline position, and timing of breaths can all change over
time.
The goal of this PhD project is to create a biomechanical digital twin of the respiratory
system, integrated with an intelligent neural state-space observer, to estimate and predict the
movement of pulmonary tumors. An offline high-fidelity finite element model (FEM) of the lung
will serve as a physics-based knowledge generator. The FEM will compute the deformation of the
lung and the trajectories of the tumor for a variety of respiratory, anatomical and pathological
situations. These simulations will generate synthetic data that can be used to learn a reduced
representation of the underlying pulmonary biomechanical dynamics. The neural state-space
model learns from the biomechanical simulations the latent dynamics that control respiratory
deformation and tumor motion.